A Robot dynamic analysis defines mathematical equations that describe the relationship between robot motion and torques. Newton-Euler and Lagrange-Euler are methods to determine the dynamic model. The performance of robot dynamic control depends on many factors, such as estimation of dynamic parameters, computation complexity, inaccuracy of the friction model parameters, simplification of the robot model, redundancy, etc. This paper implemented several Deep Learning architectures to model and establish dynamic control to overcome the problems of robustness and unmodeled dynamics. The proposed Deep Learning networks are Artificial Neural Networks, Convolutional Neural Networks, Long-short-term, Gated Recurrent units, and Bidirectional. In this context, we associate the achieved Deep learning models’ performances with specific tuning hyper-parameters such as the number of artificial neurons in hidden layers, number of hidden layers, optimization algorithms, number of epochs, attention mechanism, etc. The Bidirectional Long-short-term memory model outperformed all proposed networks. To be as close as possible to the experimental results, we add three types of noise (Uniform, Heteroscedastic, and Functional) to the training dataset to verify which of the five proposed neural networks is more effective. Furthermore, we evaluate the influence of geometric robotic complexity on the DL models’ performance. In this work, we also propose a new non-linear motion control technique. We conduct tests using numerical and deep learning methods to comprehensively verify the effectiveness and robustness of the proposed nonlinear control technique.